Papers by Caiqi Zhang
Conformity in Large Language Models (2025.acl-long)
Copied to clipboard
| Challenge: | Conformity is a form of social influence that affects the way people respond to information. |
| Approach: | They adapt psychological experiments to examine the extent of conformity in large language models. |
| Outcome: | The proposed interventions mitigate conformity by reducing the naturalness of majority tones and reducing instruction-tuned models. |
All Roads Lead to Rome: Graph-Based Confidence Estimation for Large Language Model Reasoning (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for confidence estimation are primarily designed for factual QA tasks and fail to generalize to reasoning tasks. |
| Approach: | They propose a set of training-free, graph-based confidence estimation methods tailored to reasoning tasks that exploit graph properties such as centrality, path convergence, and path weighting. |
| Outcome: | The proposed methods improve confidence estimation and performance on two downstream tasks. |
Can Large Language Models Generate High-quality Patent Claims? (2025.findings-naacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) have shown exceptional performance across various text generation tasks, but remain under-explored in the patent domain, which offers highly structured and precise language. |
| Approach: | They construct a dataset to investigate the performance of current LLMs in patent claim generation. |
| Outcome: | The proposed model outperforms state-of-the-art general LLMs in patent claim generation. |
Decoupling the Effect of Chain-of-Thought Reasoning: A Human Label Variation Perspective (2026.findings-acl)
Copied to clipboard
| Challenge: | Reasoning-tuned large language models (LLMs) with long Chain-of-Thought excel at single-answer tasks, yet their ability to model Human Label Variation remains underexplored. |
| Approach: | They conduct systematic disentanglement experiments to isolate the effect of reasoning text from intrinsic model priors on distribution-based tasks. |
| Outcome: | The proposed model improves distributional alignment, but distributional ranking is governed by model priors. |
LUQ: Long-text Uncertainty Quantification for LLMs (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing research on Uncertainty Quantification (UQ) predominantly targets short text generation, however, real-world applications often necessitate much longer responses. |
| Approach: | They propose a method that ensembles responses from multiple models and selects the response with the lowest uncertainty. |
| Outcome: | The proposed method outperforms baseline methods in correlating with the model’s factuality scores (negative coefficient of -0.85 observed for Gemini Pro). |
Do We Need Language-Specific Fact-Checking Models? The Case of Chinese (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing fact-checking models in other languages lack grounding in real-world claims . current models are constrained to a single domain, like COVID-19 . |
| Approach: | They propose a Chinese document-level evidence retriever that can be translated into Chinese . they then construct an adversarial dataset that is more robust toward biases . |
| Outcome: | The proposed method outperforms translation-based methods and multilingual language models and is more robust toward biases. |
Demystifying Multi-Agent Debate: The Role of Confidence and Diversity (2026.findings-acl)
Copied to clipboard
| Challenge: | Multi-agent debate (MAD) is widely used to improve large language models' (LLMs) reasoning and test-time scaling. |
| Approach: | They propose a diversity-aware initialisation that selects a more diverse pool of candidate answers, increasing the likelihood that a correct hypothesis is present at the start of debate. |
| Outcome: | The proposed protocol outperforms vanilla MAD and majority vote on six reasoning-oriented QA benchmarks. |
A Head to Predict and a Head to Question: Pre-trained Uncertainty Quantification Heads for Hallucination Detection in LLM Outputs (2025.emnlp-main)
Copied to clipboard
Artem Shelmanov, Ekaterina Fadeeva, Akim Tsvigun, Ivan Tsvigun, Zhuohan Xie, Igor Kiselev, Nico Daheim, Caiqi Zhang, Artem Vazhentsev, Mrinmaya Sachan, Preslav Nakov, Timothy Baldwin
| Challenge: | Uncertainty quantification (UQ) is a framework for assessing the reliability of model outputs. |
| Approach: | They introduce pre-trained UQ heads for LLMs that are highly robust and generalized to languages they were not explicitly trained on. |
| Outcome: | The pre-trained heads significantly improve their ability to capture uncertainty compared to unsupervised methods. |
TopViewRS: Vision-Language Models as Top-View Spatial Reasoners (2024.emnlp-main)
Copied to clipboard
| Challenge: | Top-view perspective is a typical way in which humans read and reason over different types of maps, but spatial reasoning capabilities of modern VLMs in this setup remain unattested and underexplored. |
| Approach: | They introduce a top-view spatial reasoning dataset and use it to evaluate VLMs across 4 perception and reasoning tasks with different levels of complexity. |
| Outcome: | The proposed model can understand and reason over spatial relations from the top view and can be controlled at different granularities of spatial reasoning. |
UNCLE: Benchmarking Uncertainty Expressions in Long-Form Generation (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing work lacks direct and fair evaluation of Large Language Models’ ability to express uncertainty effectively in long-form generation. |
| Approach: | They propose a benchmark to evaluate uncertainty expression in both long- and short-form question answering (QA) they propose prompt-based and training-based methods to improve models’ performance. |
| Outcome: | The proposed method mitigates this issue but a misalignment persists in uncertainty expression between long- and short-form generation. |
LoGU: Long-form Generation with Uncertainty Expressions (2025.acl-long)
Copied to clipboard
Ruihan Yang, Caiqi Zhang, Zhisong Zhang, Xinting Huang, Sen Yang, Nigel Collier, Dong Yu, Deqing Yang
| Challenge: | Large Language Models (LLMs) generate factually incorrect content, i.e., hallucinations, despite impressive performance. |
| Approach: | They propose a framework to enable models to express uncertainty when unsure . they propose atomic claims to refine uncertainty and refine it using supervised fine-tuning and direct preference optimization to enhance uncertainty expression. |
| Outcome: | The proposed framework significantly improves accuracy, reduces hallucinations, and maintains comprehensiveness of responses. |
Learning Action Conditions from Instructional Manuals for Instruction Understanding (2023.acl-long)
Copied to clipboard
| Challenge: | a weakly supervised task is proposed to extract mentions of preconditions and postconditions of actions from instructional manuals. |
| Approach: | They propose a task dubbed action condition inference which extracts mentions of preconditions and postconditions of actions from instructional manuals. |
| Outcome: | The proposed approach improves on the existing models, but still far behind human performance. |
LoVeC: Reinforcement Learning for Better Verbalized Confidence in Long-Form Generation (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for hallucination detection are limited to short-form question answering tasks and do not generalize well to open-ended generation. |
| Approach: | They propose a method that trains LLMs to append a numerical confidence score to each generated statement during long-form generation. |
| Outcome: | The proposed method is 20 faster than traditional self-consistency methods while achieving better calibration. |
Confidence Estimation for LLMs in Multi-turn Interactions (2026.findings-acl)
Copied to clipboard
Caiqi Zhang, Ruihan Yang, Xiaochen Zhu, Chengzu Li, Tiancheng Hu, Yijiang River Dong, Deqing Yang, Nigel Collier
| Challenge: | Despite recent progress, most prior work studies confidence in single-turn question answering. |
| Approach: | They propose a logit-based probe that measures confidence in multi-turn dialogues . they propose 'infoECE' and a "hinter-guesser" paradigm for generating controlled evaluations based on data . |
| Outcome: | The proposed framework is grounded in calibration and monotonicity of confidence as more information becomes available. |
Language is All a Graph Needs (2024.findings-eacl)
Copied to clipboard
| Challenge: | Existing work on integrating graph problems into generative language modeling framework remains limited. |
| Approach: | They propose an LLM with instructions based on natural language to perform graph tasks. |
| Outcome: | The proposed model surpasses all GNN baselines on ogbn-arxiv, Cora and PubMed datasets and sheds light on generative LLMs as new foundation model for graph machine learning. |
Lost in Embeddings: Information Loss in Vision–Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Experiments reveal connectors substantially distort the local geometry of visual representations, with k-nearest neighbors diverging by 40–60% post-projection, correlating with degradation in retrieval performance. |
| Approach: | They propose two approaches to examine and quantify information loss by analyzing latent representation space. |
| Outcome: | The proposed model improves retrieval performance by analyzing changes in k-nearest neighbor relationships between image representations before and after projection. |
Value of Information: A Framework for Human–Agent Communication (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to large language model (LLM) agents fail to account for stakes of different decisions. |
| Approach: | They propose a framework that balances task risk, query ambiguity, user effort . they use a value-of-information framework to dynamically weigh the expected utility gain . |
| Outcome: | The proposed model matches or exceeds the best manually-tuned baselines in four domains . it explicitly balances task risk, query ambiguity, and user effort . |